11,001 New Features for Statistical Machine Translation

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 218–226, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics 11,001 New Features for Statistical Machine Translation∗ David Chiang and Kevin Knight USC Information Sciences Institute 4676 Admiralty Way, Suite 1001 Marina del Rey, CA 90292 USA Wei Wang Language Weaver, Inc. 4640 Admiralty Way, Suite 1210 Marina del Rey, CA 90292 USA Abstract We use the Margin Infused Relaxed Algo- rithm of Crammer et al. to add a large num- ber of new features to two machine transla- tion systems: the Hiero hierarchical phrase- based translation system and our syntax-based translation system. On a large-scale Chinese- English translation task, we obtain statistically significant improvements of +1.5 B and +1.1 B, respectively. We analyze the im- pact of the new features and the performance of the learning algorithm. 1 Introduction What linguistic features can improve statistical ma- chine translation (MT)? This is a fundamental ques- tion for the discipline, particularly as it pertains to improving the best systems we have. Further: • Do syntax-based translation systems have unique and effective levers to pull when design- ing new features? • Can large numbers of feature weights be learned efficiently and stably on modest amounts of data? In this paper, we address these questions by exper- imenting with a large number of new features. We add more than 250 features to improve a syntax- based MT system—already the highest-scoring sin- gle system in the NIST 2008 Chinese-English common-data track—by +1.1 B. We also add more than 10,000 features to Hiero (Chiang, 2005) and obtain a +1.5 B improvement. ∗This research was supported in part by DARPA contract HR0011-06-C-0022 under subcontract to BBN Technologies. Many of the new features use syntactic informa- tion, and in particular depend on information that is available only inside a syntax-based translation model. Thus they widen the advantage that syntax- based models have over other types of models. The models are trained using the Margin Infused Relaxed Algorithm or MIRA (Crammer et al., 2006) instead of the standard minimum-error-rate training or MERT algorithm (Och, 2003). Our results add to a growing body of evidence (Watanabe et al., 2007; Chiang et al., 2008) that MIRA is preferable to MERT across languages and systems, even for very large-scale tasks. 2 Related Work The work of Och et al (2004) is perhaps the best- known study of new features and their impact on translation quality. However, it had a few shortcom- ings. First, it used the features for reranking n-best lists of translations, rather than for decoding or for- est reranking (Huang, 2008). Second, it attempted to incorporate syntax by applying off-the-shelf part-of- speech taggers and parsers to MT output, a task these tools were never designed for. By contrast, we incor- porate features directly into hierarchical and syntax- based decoders. A third difficulty with Och et al.’s study was that it used MERT, which is not an ideal vehicle for fea- ture exploration because …

Content truncated to 500 words